Getting Started

Your SystemUnknown OS • GPU not detected

Thumper-Run Documentation

Thumper-Run is a local-first platform for running AI apps on your own hardware. Install apps from the catalog, manage AI models, and opt into network features when needed. Private end-to-end encryption remains planned.

Design Goals

Local-first — everything runs on your hardware, no cloud required
Zero config — GPU detection, model discovery, and dependency management are automatic
Local control — local workflows do not require an account; sync, cloud, and telemetry are separate opt-in features
One-click install — from catalog to running app in a single click

How It Works

  Catalog          Install           Launch            Use
  ┌────────┐      ┌──────────┐     ┌─────────┐      ┌────────┐
  │ Browse │ ─→  │ Clone +  │ ─→ │ Start + │ ─→  │ WebView│
  │  apps  │      │ pip/conda│     │ health  │      │  or    │
  │        │      │ + models │     │  check  │      │ browser│
  └────────┘      └──────────┘     └─────────┘      └────────┘

What Makes It Different

Feature	Thumper-Run	Manual Install	Cloud API
GPU auto-detection	✓	✗	N/A
One-click install	✓	✗	✓
Runs offline	✓	✓	✗
Private E2EE sync (planned)	✓	✗	✗
Model management	✓	✗	N/A
Free / open source	✓	✓	✗ (per-use pricing)

Quickstart

User Quickstart

Download — grab the installer from the download page
Install an app — open the catalog, pick ComfyUI, click Install
Generate — click Launch, load the default workflow, hit Queue Prompt

Developer Quickstart

Create a minimal manifest, test it locally, and submit to the catalog:

Copy the minimal manifest below into my-app.thumper.yaml
Edit the 5 required fields (id, name, version, runtime.type, runtime.entry)
Run thumper install ./my-app.thumper.yaml to test locally

yaml

thumper: "1.0"
id: my-app
name: My App
version: "0.1.0"
runtime:
  type: python
  entry: main.py

Installation

Download the desktop app from the download page. No account required — just install and run.

Per-Platform Instructions

bash

# AppImage — download and run directly
chmod +x thumper-run.AppImage
./thumper-run.AppImage
 
# Debian / Ubuntu
sudo dpkg -i thumper-run.deb
thumper-run

Requires: glibc 2.31+, WebKitGTK 4.1+ (installed automatically on most distros)

System Requirements

	Minimum	Recommended
RAM	8 GB	16 GB
Disk	10 GB	50 GB+ (for models)
GPU	None (CPU mode)	8 GB+ VRAM
OS	Linux x86_64, macOS (Intel/Apple Silicon), Windows x86_64	—

GPU Auto-Detection

On first launch, the platform detects your GPU and configures the environment:

GPU	Backend	PyTorch Index	Notes
NVIDIA	CUDA	cu121 / cu124	Auto-matched to driver version
AMD (discrete)	ROCm	rocm6.2	Requires ROCm-capable GPU (gfx900+)
AMD APU (iGPU)	ROCm	TheRock nightly	BF16 only; FP16 causes NaN on gfx1150
Intel Arc	oneAPI	xpu	Level Zero required
Apple Silicon	MPS	cpu (MPS via PyTorch)	Shared memory; never use --lowvram
None / CPU only	CPU	cpu	Works for LLMs (GGUF), slower for image gen

AMD APU (gfx1150): FP16 causes NaN in DiT attention. Always use BF16. See Troubleshooting for details.

Apple Silicon: Shared memory — never use --lowvram. It causes 5% GPU utilization by shuffling data through the same physical RAM.

GPU Compatibility Matrix

GPU	Status	Backend	Notes
NVIDIA RTX 30xx/40xx	Fully supported	CUDA 12.1+	Best compatibility and performance
NVIDIA RTX 20xx	Fully supported	CUDA 12.1+	Turing architecture, good performance
NVIDIA GTX 10xx	Limited	CUDA 12.1+	Works but low VRAM (6–8 GB max)
AMD RX 7000	Fully supported	ROCm 6.2	RDNA 3, excellent ROCm support
AMD RX 6000	Limited	ROCm 6.2	RDNA 2, some models may need workarounds
AMD Ryzen AI (gfx1150)	Limited	TheRock nightly	BF16 only, shared memory, MIOpen warmup
Intel Arc A770/A750	Limited	oneAPI / XPU	Level Zero required, some models unsupported
Apple M1/M2/M3/M4	Fully supported	MPS	Shared memory, never use --lowvram
CPU only	Fully supported	CPU	Best for GGUF LLMs, slow for image gen
Older NVIDIA (pre-Pascal)	Unsupported	—	Use CPU mode instead

First Launch

The first time you open Thumper-Run you'll see the app catalog. No login is needed — everything works anonymously by default. You can optionally:

Add a recovery phrase — enables multi-device sync and data recovery if you lose your device
Sign in with Google/GitHub — links to your identity for easier recovery, optional
Skip both — use fully anonymously, device-local only

Learn More

What is Thumper-Run?

Platform overview and key concepts

Tutorials

Hands-on guides from beginner to intermediate

Troubleshooting

Symptom-based diagnosis and benchmarks

File System Layout

Where files are stored on each platform